You Are What You Annotate: Towards Better Models through Annotator Representations
share
Annotator disagreement is ubiquitous in natural language processing (NLP) tasks. There are multiple reasons for such disagreements, including the subjectivity of the task, difficult cases, unclear guidelines, and so on. Rather than simply aggregating labels to obtain data annotations, we instead propose to explicitly account for the annotator idiosyncrasies and leverage them in the modeling process. We create representations for the annotators (annotator embeddings) and their annotations (annotation embeddings) with learnable matrices associated with each. Our approach significantly improves model performance on various NLP benchmarks by adding fewer than 1 parameters. By capturing the unique tendencies and subjectivity of individual annotators, our embeddings help democratize AI and ensure that AI models are inclusive of diverse viewpoints.
READ FULL TEXT